Abstract

This paper describes the systems submitted by IIT (BHU), Varanasi/IIIT Hyderabad (IITBHU–IIITH) for Task 1 of CoNLL– SIGMORPHON 2018 Shared Task on Universal Morphological Reinflection (Cotterell et al., 2018). The task is to generate the inflected form given a lemma and set of morphological features. The systems are evaluated on over 100 distinct languages and three different resource settings (low, medium and high). We formulate the task as a sequence to sequence learning problem. As most of the characters in inflected form are copied from the lemma, we use Pointer-Generator Network (See et al., 2017) which makes it easier for the system to copy characters from the lemma. PointerGenerator Network also helps in dealing with out-of-vocabulary characters during inference. Our best performing system stood 4th among 28 systems, 3rd among 23 systems and 4th among 23 systems for the low, medium and high resource setting respectively

Abstract

Humans can learn multiple languages. If they know a fact in one language, they can answer a question in another language they understand. They can also answer Code-mix (CM) questions: questions which contain both languages. This behavior is attributed to the unique learning ability of humans. Our task aims to study if machines can achieve this. We demonstrate how effectively a machine can answer CM questions. In this work, we adopt a two phase approach: candidate generation and candidate re-ranking to answer questions. We propose a Triplet-Siamese-Hybrid CNN (TSHCNN) to re-rank candidate answers. We show experiments on the SimpleQuestions dataset. Our network is trained only on English questions provided in this dataset and noisy Hindi translations of these questions and can answer English-Hindi CM questions effectively without the need of translation into English. Back-transliterated CM questions outperform their lexical and sentence level translated counterparts by 5% & 35% in accuracy respectively, highlighting the efficacy of our approach in a resource constrained setting.

Abstract

In this paper, we describe Transzaar -an AI powered tool that offers computer aided translation (CAT) functionality: pre-translation analysis,post-editing machine translated content, translation prediction, text aligning, extensive logging and integration with several machine translation (MT) systems. Transzaar aids a human translator to perform various language processing tasks, viz., Translation, Transliteration, Localization, and other kinds of Text Analysis tasks. Using Transzaar, human translators can post-edit the machine translated content, to improve fluency and accuracy of the translated content to match the naturalness of human translation while delivering with better turn-around time.Transzaar aids the process of post-editing, thereby increasing the productivity of human translators by 2-3 folds within couple of months of usage for certain language pairs. It collects feedback continuously, which helps the MT system to further learn and improve periodically, with the additional new generated data-set

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

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

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

Keywords of a document are a representative of its content, and it helps to have meaningful words to facilitate search and organization of documents. Hence, finding methods that can automatically identify keywords in a document is very important as manual processes for this is very cumbersome and error-prone. If this task is accomplished for song lyrics, it has varied applications such as recommendation systems and digital music library management. This work proposes and compares methods to identify keywords from lyrics of Bollywood songs. We use a collection of lyrics of 1055 Bollywood songs, all written in the Devanagari script. Experiments include looking at the spatial distribution of the terms, their occurrence in a certain context or position, and using WordNet to generate keywords not present in the document. Validation was done by human annotators by providing a score to each method based on the results obtained on a subset of the data. We also used Latent Dirichlet Allocation and Latent Semantic Indexing to validate the results, as further explained in the paper.

Abstract

Many methods have been developed for various sentence classification tasks for English, which usually exploit linguistic resources like parsers or rely on the large amount of annotated or unannotated data, making it difficult to adapt them to other languages. In this paper, we present an evaluation of popular deep learning methods for sentence classification on the morphologically rich Indian languages, specifically, Hindi and Telugu. For this purpose, we also created a question classification dataset for Hindi, by translating the TREC-UIUC dataset. We show that character based input can enhance the performance of current classification systems for morphologically rich languages. Finally, we show that our multiInput-CNN variant is able to perform better than our baselines in two out of three tasks in Hindi and Telugu, while giving comparable results for others.

Abstract

The paper presents the development of a phonetically balanced read speech corpus of code-mixed Hindi-English. Phonetic balance in the corpus has been created by selecting sentences that contained triphones lower in frequency than a predefined threshold. The assumption with a compulsory inclusion of such rare units was that the high frequency triphones will inevitably be included. Using this metric, the Pearson’s correlation coefficient of the phonetically balanced corpus with a large code-mixed reference corpus was recorded to be 0.996. The data for corpus creation has been extracted from selected sections of Hindi newspapers. These sections contain frequent English insertions in a matrix of Hindi sentence. Statistics on the phone and triphone distribution have been presented, to graphically display the phonetic likeness between the reference corpus and the corpus sampled through our method.

Abstract

In order for our computer systems to be more human-like, with a higher emotional quotient, they need to be able to process and understand intrinsic human language phenomena like humour. In this paper, we consider a subtype of humour - puns, which are a common type of wordplay-based jokes. In particular, we consider code-mixed puns which have become increasingly mainstream on social media, in informal conversations and advertisements and aim to build a system which can automatically identify the pun location and recover the target of such puns. We first study and classify code-mixed puns into two categories namely intra-sentential and intra-word, and then propose a four-step algorithm to recover the pun targets for puns belonging to the intra-sentential category. Our algorithm uses language models, and phonetic similarity-based features to get the desired results. We test our approach on a small set of code-mixed punning advertisements, and observe that our system is successfully able to recover the targets for 67% of the puns.