China
Africa
Explore chapters and articles related to this topic
A Computational Linguistic Approach to Modelling the Dynamics of Design Processes
Published in Bo T. Christensen, Linden J. Ball, Kim Halskov, Analysing Design Thinking: Studies of Cross-Cultural Co-Creation, 2017
Joel Chan, Christian D. Schunn
Our evaluation criterion was the real world coherence of the topics that resulted. We used CV, an automated coherence measure developed by Röder, Both, and Hinneburg (2015). This measure formalizes the intuition that the set of words that compose a topic (i.e., the words that have high probability given a particular topic) should tend to co-occur more often in the “real world”: for example, the topic {football, pitch, boots, jersey} is more coherent than the topic {football, kitchen, computer, dress} because the words in the first topic tend to co-occur more often than the words in the second topic. In the case of CV, the “real world” is defined as co-occurrence patterns of words in the English Wikipedia. Although the mathematical details of the measure are out of the scope of this paper, the interested reader is referred to Röder et al. (2015) for more details. Early use of LDA used perplexity to evaluate topic models. Perplexity measures the model’s ability to predict what words would show up in a new set of documents that the model has not yet seen. However, topic coherence is the new standard for evaluating topic models, since Chang and colleagues (2009) showed that the perplexity measure is very often uncorrelated with human-judged topic interpretability.
Speech Recognition
Published in Sadaoki Furui, Digital Speech Processing, Synthesis, and Recognition, 2018
Three measures for representing the syntactic complexity of recognizing tasks have thus far been proposed to facilitate the evaluation of the difficulty of speech recognition tasks. The average branching factor indicates the average number of words which can be predicated, that is, the words that can follow at each position of syntactic analysis (Goodman, 1976). Equivalent vocabulary size is a modification of the average branching factor in which the acoustic similarity between words is taken into consideration (Goodman, 1976). Finally, perplexity is defined by 2H, where H is the entropy of a word string in sentence speech (Bahl et al., 1983). Entropy H is given by equation H=-∑(w1w2⋯wk)1kP(w1w2⋯wk)log2P(w1w2⋯wk) $$ H = - \sum\limits_{{(w_{1} w_{2} \cdots w_{k} )}} {\frac{1}{k}P(w_{1} w_{2} \cdots w_{k} ){\text{log}}_{2} P(w_{1} w_{2} \cdots w_{k} )} $$
Transformers: Basics and Introduction
Published in Uday Kamath, Kenneth L. Graham, Wael Emara, Transformers for Machine Learning, 2022
Uday Kamath, Kenneth L. Graham, Wael Emara
Perplexity is an intrinsic evaluation method often used for measuring the performance of tasks such as language modeling and machine translation [42]. In Fig. 2.16, we show plots of perplexity and training/validation loss for the attention model, measured per epoch during the training process. It is interesting to note that the validation loss starts flattening out around epoch 14 and training loss further reduces, thus indicating overfitting. The best model for both approaches are chosen based on the best fit on the validation data.
Data-driven review of blockchain applications in supply chain management: key research themes and future directions
Published in International Journal of Production Research, 2023
Truong Van Nguyen, Hiep Cong Pham, Minh Nhat Nguyen, Li Zhou, Mohammadreza Akbari
Another issue of LDA is that it requires users to predefine the number of latent topics before clustering. A common way to select the optimal value of in LDA is using the perplexity score of the holdout set to evaluate the clustering performance of the trained LDA model (Blei, Ng, and Jordan 2003). The perplexity score can be measured using Equation (3). Generally, the lower perplexity score indicates better model performance. Prior to exploring the topics/themes using LDA, it is important to convert the corpus of document texts into a mathematical object called document term matrix (DTM), so that quantitative techniques such as clustering can be applied. Typically, DTM describes the occurrence of all terms (i.e. words) in each document. Since texts can be highly unstructured and noisy, we apply some text processing techniques to remove terms that do not contain semantic information before creating the DTM. These techniques include tokenising sentences into terms, lower case transformation, non-letter removal, English stop word removal and truncating terms to their base form using stemming (i.e. the process to reduce the inflected words into their root form). After this text processing procedure, the DTM only contains keywords with semantic information to be fed into the LDA mechanism described earlier.
What we talk about when we talk about EEMs: using text mining and topic modeling to understand building energy efficiency measures (1836-RP)
Published in Science and Technology for the Built Environment, 2023
Apoorv Khanuja, Amanda L. Webb
Even though topic modeling is an unsupervised algorithm, the expected number of topics (k) still needs to be specified. If the value of k is too low, the LDA model will be too coarse to differentiate between topics. However, if the value of k is too high, it will make the model too complex and granular. For this analysis, the perplexity values for topic models from k = 2 to k = 12 topics were calculated using the topicmodels R package (Grün and Hornik 2011). Perplexity is a statistical measure of how well a probability model predicts a sample, with low values meaning that the model is a better predictor (Blei, Ng, and Jordan 2003; Zhao et al. 2015). Six topics were selected for this analysis based on a combination of diminishing returns in the perplexity analysis curve and keeping the number of topics relatively small.
Understanding water disputes in Chile with text and data mining tools
Published in Water International, 2019
Mauricio Herrera, Cristian Candia, Diego Rivera, Douglas Aitken, Daniel Brieba, Camila Boettiger, Guillermo Donoso, Alex Godoy-Faúndez
For LDA, it is necessary to specify in advance the number of topics in the underlying topic structure. We used perplexity (Berthard, Ghosh, Martin, & Summer, 2009) to determine the optimal number of themes. Perplexity is a statistical measure of how well a probability model predicts a sample. As applied to LDA, for a given value of the number of topics, the parameters for the LDA model are estimated. Then, given the theoretical word distributions represented by the topics, the result is compared to the actual theme mixtures, or the distribution of words in the documents. This statistic is somewhat meaningless on its own. The benefit comes from comparing perplexity across different models with varying numbers of topics. The model with the lowest perplexity is generally considered the best.