Context Aware Conversational Understanding for Intelligent Agents With a Screen

Authors: Vishal Naik, Angeliki Metallinou, Rahul Goel

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show that this approach outperforms a rule-based alternative, and can be extended in a straightforward manner to new contextual use cases. We perform detailed evaluation of contextual and non-contextual use cases and show that our system displays accurate contextual behavior without degrading the performance of noncontextual user requests.
Researcher Affiliation Collaboration Vishal Ishwar Naik, 1 Arizona State University, 2 Amazon Alexa Machine Learning vnaik1@asu.edu, {ametalli, goerahul }@amazon.com
Pseudocode No The paper describes models and architectures with mathematical notations and figures but does not include explicit pseudocode blocks or algorithm listings.
Open Source Code No The paper mentions 'word2vec pre-trained embeddings of size E=300 downloaded from (Mikolov )' with a URL to a Google Code Archive for word2vec, but this is a third-party resource they used, not their own source code for the methodology described in the paper. There is no explicit statement or link for their own code.
Open Datasets No We use a large set of non contextual utterances from user interactions with Alexa on devices without a screen, and a smaller dataset of contextual utterances from user interactions where a screen is available. ... The dataset used for these experiments is a fraction of our production data and covers a range of domain functionality including music, books, movies and showtimes, videos, calendar events, local search, shopping, and general commands.
Dataset Splits Yes We used a train-dev split of 70%-30%, where the dev set was used for optimizing the deep learning models as well as tuning the reranker parameters α, β.
Hardware Specification No The paper mentions 'availability of fast GPU computing resources' in the related work section but does not specify the particular hardware (e.g., GPU model, CPU type) used for their experiments.
Software Dependencies No The paper mentions deep learning models (LSTMs, CNNs) and word2vec embeddings but does not provide specific software dependencies with version numbers (e.g., Python, TensorFlow/PyTorch versions, specific library versions).
Experiment Setup Yes All our deep learning models were trained end-to-end using stochastic gradient descent. Models were regularized using dropout and L2. We used word2vec pre-trained embeddings of size E=300 downloaded from (Mikolov ). For the context-Bi LSTM(keys,values) model we use a hidden layer of size H between the context value feature vector and the Bi LSTM, see Fig. 1b. Empirically, we chose H = 200 while the input value vector is of size N E = 5 300. Value inputs at each of the 5 positions are shuffled between training epochs, as described in 5.2, which led to better performance. For the context-CNN+Bi LSTM(keys,values), we used F = 100 filters for the convolutional layer. We also found that adding a Re Lu non-linearity after the max pooling operation performs slightly better. For the non-context Bi LSTM + contextual Reranker, the reranker parameters α and β were chosen based on grid search in [0, 1], to optimize the dev set performance for slots and intents.