Abstract

Extractive summarization methods operate by ranking and selecting the sentences which best encapsulate the theme of a given document. They do not fare wellin domains like fictional narratives where there is no central theme and core in-formation is not encapsulated by a small set of sentences. For the purpose ofreducing the size of the document while conveying the idea expressed by each sentence, we need more sentence specific methods. Telegraphic summarization,which selects short segments across several sentences, is better suited for such domains. Telegraphic summarization captures the plot better by retaining shorter versions of each sentence while not really concerning itself with grammatically linking these segments. In this paper, we propose an unsupervised deep learning network (NUTS) to generate telegraphic summaries. We use multiple encoder-decoder networks and learn to drop portions of the text that are inferable from thechosen segments. The model is agnostic to both sentence length and style. We demonstrate that the summaries produced by our model show significant quanti-tative and qualitative improvement over those produced by existing methods and baselines.

Abstract

n the past few years, bully and aggressive posts on social media have grown signifi-cantly, causing serious consequences for vic-tims/users of all demographics. Majority of the work in this field has been done for En-glish only. In this paper, we introduce a deep learning based classification system for Face-book posts and comments of Hindi-English Code-Mixed text to detect the aggressive be-haviour of/towards users. Our work focuses ontext from users majorly in the Indian Subcon-tinent. The dataset that we used for our mod-els is provided byTRAC-11in their shared task. Our classification model assigns each Facebook post/comment to one of the threepredefined categories: “Overtly Aggressive”,“Covertly Aggressive” and “Non-Aggressive”.We experimented with 6 classification mod-els and our CNN model on a 10 K-fold cross-validation gave the best result with the predic-tion accuracy of 73.2%

Abstract

In a multilingual or sociolingual configuration Intra-sentential Code Switching (ICS) or Code Mixing (CM) is frequently observed nowadays. In the world most of the people know more than one language. The CM usage is especially apparent in social media platforms. Moreover, ICS is particularly significant in the context of technology, health and law where conveying the upcoming developments are difficult in ones native language. In applications like dialog systems, machine translation, semantic parsing, shallow parsing, etc. CM and Code Switching pose serious challenges. To do any further advancement in code-mixed data, the necessary step is Language Identification. So, in this paper we present a study of various models - Nave Bayes Classifier, Random Forest Classifier, Conditional Random Field (CRF) and Hidden Markov Model (HMM) for Language Identification in English - Telugu Code Mixed Data. Considering the paucity of resources in code mixed languages, we proposed CRF model and HMM model for word level language identification. Our best performing system is CRF-based with an f1-score of 0.91.

Abstract

In this paper, we model the similar question retrieval task as a binary classification problem. We propose a novel approach of “1D-Siamese LSTM for cQA (1D-SLcQA)” to find the semantic similarity between a new question and existing question(s). In 1D-SLcQA, we use a combination of twin LSTM networks and a contrastive loss function to effectively memorize the long term dependencies i.e., capture semantic similarity even when the length of the answers/questions is very large (200 words). The similarity of the questions is modeled using a single network with (1D) (feature) convolution between feature vectors learned from twin LSTM layers. Experiments on large scale real world Yahoo Answers dataset show that 1D-SLcQA outperform the state of the art approach of Siamese cQA approach (SCQA).

Abstract

Movies are one of the most prominent means of entertainment. The widespread use of the Internet in recent times has led to large volumes of data related to movies being generated and shared online. People often prefer to express their views online in English as compared to other local languages. This leaves us with a very little amount of data in languages apart from English to work on. To overcome this, we created the Multi-Language Movie Review Dataset (MLMRD). The dataset consists of genre, rating, and synopsis of a movie across multiple languages, namely Hindi, Telugu, Tamil, Malayalam, Korean, French, and Japanese. The genre of a movie can be identified by its synopsis. Though the rating of a movie may depend on multiple factors like the performance of actors, screenplay, direction etc but in most of the cases, synopsis plays a crucial role in the movie rating. In this work, we provide various model architectures that can be used to predict the genre and the rating of a movie across various languages present in our dataset based on the synopsis

Abstract

While extractive summarization is a well studied problem, it is far from solved. In recent years a large number of interesting and complex models have been used to achieve significant improvements in performance. This can easily be attributed to Deep Learning models and dense vector representations but the performance gain comes with the cost of computational and representational complexity. In this work, we present a simple, yet effective approach for extractive summarization of news articles. In line with many recent works in this area we propose an encoder-decoder architecture with a simple bag of word encoder for sentences followed by an attention based decoder for relevant sentence selection. Our model is trained end-to-end and its performance is comparable to the state-of-the-art models while being simpler both in terms of the number of parameters (significantly lesser) as well as the representational complexity.

Abstract

When a human interacts with an information retrieval chat bot, he/she can ask multiple questions at the same time. Current question answering systems can’t handle this scenario effectively. In this paper we propose an approach to identify question spans in a given utterance, by posing this as a sequence labeling problem. The model is trained and evaluated over 4 different freely available datasets. To get a comprehensive coverage of the compound question scenarios, we also synthesize a dataset based on the natural question combination patterns. We exhibit improvement in the performance of the DrQA system when it encounters compound questions which suggests that this approach is vital for real-time human-chatbot interaction.

Abstract

Humor is an essential characteristic of language. It has been a topic of research in linguistics and philosophy from historical times. In computer science, computational humor, as a part of Natural Language Processing, is a growing area of research. Social Media is rapidly growing as a platform for communication but processing of social media, owing to its semantic perplexity, is still a challenge. These two facts lead us to present a novel dataset for humor classification which captures diversity in humor on web resources. The large size of this dataset is to meet the data requirements for modern machine learning algorithms. This paper also deals with creating a model for detecting and analyzing humor in social media text extracted from eclectic sources on the Internet.

Abstract

Resource poor languages, like Odia, inherently lack the necessary resources and tools for the task of sentiment analysis to give promising results. With more user-generated raw data readily available today, it is of prime importance to have annotated corpora from various domains. This paper is a first attempt towards building an annotated corpus of Odia poetry with sentiment labels. The annotated corpus is further used for sentiment classification using machine learning techniques in order to establish a baseline. Stylistic variations and structural differences between poetic and non-poetic texts make the task of sentiment classification challenging for the former. Using the annotated corpus of poems, we obtained comparable accuracies across various classification models. Linear-SVM outperformed other classifiers with a macro F1- Score of 0.68. The annotated corpus contains a total of 730 Odia Poems of various genres with a vocabulary of more than 23k words. Fleiss Kappa score of 0.83 was obtained which corresponds to near perfect agreement among the annotators.

Abstract

Sentiment analysis research has predominantly been on English texts. There exists many sentiment resources for English but very less exist for other languages. To improve sentiment analysis in a low resource language, sentiment labeled corpora are translated from English into the focus language and use them as additional resources for sentiment analysis research in the focus language [3]. But when text is translated from one language into another, sentiment is preserved to varying degrees. In this paper, we use product and book reviews in English as stand-in for source language text and determine loss in sentiment and sentiment predictability when they are translated into Telugu (a low resource South Asian language), manually and automatically. For this purpose, we use manually and automatically determined sentiment labels of the English text as a benchmark. We show that sentiment analysis of Telugu manual translations of English text produces competitive results w.r.t English sentiment analysis. We discover that even though machine translation ignificantly reduces the human ability to recover sentiment, automatic sentiment systems are still able to capture sentiment information from the translations in certain cases. In the process, we created a Telugu-English parallel corpus that is independently annotated for sentiment using a 5-value scale by Telugu and English speakers. We also created a Telugu lexicon annotated at both sentiment and emphasis level.