Abstract

We investigate the problem of parsing conversational data of morphologically-rich languages such as Hindi where argument scrambling occurs frequently. We evaluate a state-of-the-art non-linear transition-based parsing system on a new dataset containing 506 dependency trees for sentences from Bollywood (Hindi) movie scripts and Twitter posts of Hindi monolingual speakers. We show that a dependency parser trained on a newswire treebank is strongly biased towards the canonical structures and degrades when applied to conversational data. Inspired by Transformational Generative Grammar (Chomsky, 1965), we mitigate the sampling bias by generating all theoretically possible alternative word orders of a clause from the existing (kernel) structures in the treebank. Training our parser on canonical and transformed structures improves performance on conversational data by around 9% LAS over the baseline newswire parser.

Abstract

In this paper, we report our participation in the Task 2: Triple Scoring of WSDM Cup challenge 2017. In this task, we were provided with triples of "type-like" relations which were given human-annotated relevance scores ranging from 0 to 7, with 7 being the "most relevant" and 0 being the "least relevant". The task focuses on two such relations: profession and nationality. We built a system which could automatically predict the relevance scores for unseen triples. Our model is primarily a supervised machine learning based one in which we use well-designed features which are used to a make a Logistic Ordinal Regression based classification model. The proposed system achieves an overall accuracy score of 0.73 and Kendall’s tau score of 0.36.

Abstract

This paper presents the ongoing process in the design of the first phase of the phonetically balanced codemixed corpus of Hindi-English speech (PBCM-Phase I). The reference corpus is a large code-mixed (LCM) newspaper corpus selected from the sections that contain frequent English insertions in a matrix of Hindi sentence. From a phonetically transcribed corpus, compulsory inclusion of lowest frequency triphones has been ensured, with the assumption that high frequency phones may automatically be included. A high correlation of 0.81 with the representative large corpus has been observed. A small scale speech corpus of 5.6 hours has been collected, by the contribution of 4 volunteer native Hindi speakers. The recording has been conducted in a professional recording studio environment. As a second contribution, this paper also presents a baseline recognition system with pooled monolingual and codemixed speech datasets as training and testing environments.

Abstract

Trivia is any fact about an entity which is interesting due to its unusualness, uniqueness or unexpectedness. Trivia could be successfully employed to promote user engagement in various product experiences featuring the given entity. A Knowledge Graph (KG) is a semantic network which encodes various facts about entities and their relationships. In this paper, we propose a novel approach called DBpedia Trivia Miner (DTM) to automatically mine trivia for entities of a given domain in KGs. The essence of DTM lies in learning an Interestingness Model (IM), for a given domain, from human annotated training data provided in the form of interesting facts from the KG. The IM thus learnt is applied to extract trivia for other entities of the same domain in the KG. We propose two different approaches for learning the IM - a) A Convolutional Neural Network (CNN) based approach and b) Fusion Based CNN (F-CNN) approach which combines both hand-crafted and CNN features. Experiments across two different domains - Bollywood Actors and Music Artists reveal that CNN automatically learns features which are relevant to the task and shows competitive performance relative to hand-crafted feature based baselines whereas FCNN significantly improves the performance over the baseline approaches which use hand-crafted features alone. Overall, DTM achieves an F1 score of 0.81 and 0.65 in Bollywood Actors and Music Artists domains respectively.

Abstract

The paper presents an automatic speech recognition (ASR) system for code-mixed read speech in Hindi-English, developed upon the extrapolation of monolingual training resources. A monolingual Hindi acoustic model, mixed with code-mixed speech data has been implemented to train a neural network based speech recognition framework. The testing corpus also follows a similar structure: containing data from both monolingual and code-mixed speech. The shared phonetic transcription, captured in WX notation has been exploited to harness the commonality between the pooled phonesets of Hindi and English. The experiments have been conducted in two separate formulations of a trigram based language model 1) In the first experiment, the language model contains no out-of-vocabulary words, as the test utterances are included in the training of the language model. The word error rate in this case has been obtained to be 10.63 %. 2) In the second experiment, the testing utterances have been excluded from the training language model. The word error rate in this case has been obtained to be 41.66 %.

Abstract

Code-Mixing (CM) is a natural phenomenon observed in many multilingual societies and is becoming the preferred medium of expression and communication in online and social media fora. In spite of this, current Question Answering (QA) systems do not support CM and are only designed to work with a single interaction language. This assumption makes it inconvenient for multi-lingual users to interact naturally with the QA system especially in scenarios where they do not know the right word in the target language. In this paper, we present WebShodh - an end-end web-based Factoid QA system for CM languages. We demonstrate our system with two CM language pairs: Hinglish (Matrix language: Hindi, Embedded language: English) and Tenglish (Matrix language: Telugu, Embedded language: English). Lack of language resources such as annotated corpora, POS taggers or parsers for CM languages poses a huge challenge for automated processing and analysis. In view of this resource scarcity, we only assume the existence of bi-lingual dictionaries from the matrix languages to English and use it for lexically translating the question into English. Later, we use this loosely translated question for our downstream analysis such as Answer Type(AType) prediction, answer retrieval and ranking. Evaluation of our system reveals that we achieve an MRR of 0.37 and 0.32 for Hinglish and Tenglish respectively. We hosted this system online and plan to leverage it for collecting more CM questions and answers data for further improvement.

Abstract

This paper presents DILTON, a system which solves simple arithmetic word problems. DILTON first predicts the operation that is to be performed (’-’,’+’,’*’,’/’) through a deep neural network based model and then uses it to generate the answer. DILTON divides the question into two parts - world state and query as shown in Figure 1. The world state and the query are processed separately in two different networks and finally the networks are merged to predict the final operation. DILTON learns to predict operations with 8.81 % in a corpus of primary school questions. With simple similarity between the contexts of quantities appearing in the problem and the question text, we are able to identify 92.25 % of relevant quantities and solve 81% of the questions. Our code and data is publicly available

Abstract

Word embeddings are being used for several linguistic problems and NLP tasks. Improvements in solutions to such problems are great because of the recent breakthroughs in vector representation of words and research in vector space models. However, vector embeddings of phrases keeping semantics intact with words has been challenging. We propose a novel methodology using Siamese deep neural networks to embed multi-word units and fine-tune the current state-of-the-art word embeddings keeping both in the same vector space. We show several semantic relations between words and phrases using the embeddings generated by our system and evaluate that the similarity of words and their corresponding paraphrases are maximized using the modified embeddings.

Abstract

Transfer learning algorithms can be used when sufficient amount of training data is available in the source domain and limited training data is available in the target domain. The transfer of knowledge from one domain to another requires similarity between two domains. In many resource-poor languages, it is rare to find labeled training data in both the source and target domains. Active learning algorithms, which query more labels from an oracle, can be used effectively in training the source domain when an oracle is available in the source domain but not available in the target domain. Active learning strategies are subjective as they are designed by humans. It can be time consuming to design a strategy and it can vary from one human to other. To tackle all these problems, we design a learning algorithm that connects transfer learning and active learning with the well-known multi-armed bandit problem by querying …

Abstract

The furtherance of social media led people to share the reviews in various ways such as video, audio and text. Recently, the performance of sentiment classification is achieved success using neural networks. In this paper, neural network approach is presented to detect the sentiment from audio and text models. For audio, features like Mel Frequency Cepstral Coefficients (MFCC) are used to build Deep Neural Network (DNN) and Deep Neural Network Attention Mechanism (DNNAM) classifiers. From the results, it is noticed that DNNAM gives better results compared to DNN because the DNN is a frame based one where as the DNNAM is an utterance level classification thereby efficiently use the context. Additionally, textual features are extracted from the transcript of the audio input using Word2vec model. Support Vector Machine (SVM) and Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) classifiers are used to develop a sentiment model. From the experiments it is noticed the LSTM-RNN outperforms the SVM as the LSTM-RNN is able to memorize long temporal context. The performance is also significantly improved by combining both the audio and text modalities.