Disputes between two parties occur in almost all domains such as taxation, insurance, banking, healthcare, etc. The disputes are generally resolved in a specific forum (e.g., consumer court) where facts are presented, points of disagreement are discussed, arguments as well as specific demands of the parties are heard, and finally a human judge resolves the dispute by often favouring one of the two parties. In this paper, we explore the use of large language models (LLMs) as assistants for the human judge to resolve such disputes, as part of our DRAssist system. We focus on disputes from two specific domains – automobile insurance and domain name disputes. DRAssist identifies certain key structural elements (e.g., facts, aspects or disagreement, arguments) of the disputes and summarizes the unstructured dispute descriptions to produce a structured summary for each dispute. We then explore multiple prompting strategies with multiple LLMs for their ability to assist in resolving the disputes in these domains. In DRAssist, these LLMs are prompted to produce the resolution output at three different levels – (i) identifying an overall stronger party in a dispute, (ii) decide whether each specific demand of each contesting party can be accepted or not, (iii) evaluate whether each argument by each contesting party is strong or weak. We evaluate the performance of LLMs on all these tasks by comparing them with relevant baselines using suitable evaluation metrics.

In this paper, we focus on automatic extraction of statute facets from legal statutes such as Act documents. We define statute facets to be key specific aspects of a statute which can potentially be used in legal arguments. For example, Section 25F of the Industrial Disputes Act (India) contains statute facets such as workman, employer, retrenchment of workmen, continuous service for not less than one year, etc. Such statute facets are often used by lawyers as part of their argumentation and also by judges for deciding on a case. In this paper, we propose a weakly supervised technique for extracting such statute facets from legal text. We use dependency tree structure to extract candidate statute facets and use BM25 ranking function to determine statute-specificity of these candidates. We propose a set of facet types which enable us to realize the definition of statute facets in a more computational way. We use recent deep learning models in a few-shot setting to predict an appropriate facet type for each candidate. Only those candidates with high statute-specificity and for which a facet type is predicted with high confidence, are selected as acceptable statute facets. We evaluate the extracted statute facets through both direct and indirect evaluation as well as conduct a user-study to get validation and feedback from lawyers.

Argumentation mining is a growing area of research and has several interesting practical applications of mining legal arguments. Support and Attack relations are the backbone of any legal argument. However, there is no publicly available dataset of these relations in the context of legal arguments expressed in court judgements. In this paper, we focus on automatically constructing such a dataset of Support and Attack relations between sentences in a court judgment with reasonable accuracy. We propose three sets of rules based on linguistic knowledge and distant supervision to identify such relations from Indian Supreme Court judgments. The first rule set is based on multiple discourse connectors, the second rule set is based on common semantic structures between argumentative sentences in a close neighbourhood, and the third rule set uses the information about the source of the argument. We also explore a BERT-based sentence pair classification model which is trained on this dataset. We release the dataset of 20506 sentence pairs - 10746 Support (precision 77.3%) and 9760 Attack (precision 65.8%). We believe that this dataset and the ideas explored in designing the linguistic rules and will boost the argumentation mining research for legal arguments.

Soham Datta, Prabir Mallick, Sangameshwar Patil, Indrajit Bhattacharya, Girish Palshikar. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

The past decade has witnessed a rapid growth in the use of online social networks, such as Twitter, by individuals as well as communities for fast dissemination of information. However, as useful as this information might be, it also prospects the rapid spread of rumors. This phenomenon of rumor proliferation has persuaded researchers to work in the field of rumor detection by using the temporal, textual and author related features. From a sociological perspective, rumors are generally seen to be targeted at some important entities such as well-known politicians, actors, places etc. Some rumors are aimed at creating social unrest by spreading information about eye-catching events like highjacking, bomb attack etc. The existing works in rumor detection do not seem to harness this peculiar characteristic of rumor.This research proposes a novel approach to rumor detection by using entity recognition on the post text along with other features that imply reliability and consistency. Our hypothesis defines four different scores namely, i) Famousness score, ii) Rareness score, iii) Reliability score and iv) Consistency score. Experimental results show that the proposed features not only increase the performance of our model, but also out-perform the baseline approaches in terms of F1 score.

Witness testimonies are important constituents of a court case description and play a significant role in the final decision. We propose two techniques to identify sentences representing witness testimonies. The first technique employs linguistic rules whereas the second technique applies distant supervision where training set is constructed automatically using the output of the first technique. We then represent the identified witness testimonies in a more meaningful structure – event verb (predicate) along with its arguments corresponding to semantic roles A0 and A1 [1]. We demonstrate effectiveness of such representation in retrieving semantically similar prior relevant cases. To the best of our knowledge, this is the first paper to apply NLP techniques to extract witness information from court judgements and use it for retrieving prior court cases.

This paper explains how the traditional anomaly detection techniques lack in exploring various aspects and perspectives of data analysis and lacks in understanding the nature of anomalies. This paper explains the importance and need of developing anomaly detection techniques considering the domain knowledge of the field where the anomaly detection technique is to be applied. Here we have considered the notion of causality(Granger's Causality[10]) in data to perform anomaly detection. This paper states some examples supporting the need of domain specific anomaly detection techniques. It also explains a few aspects which have not been considered in anomaly detection techniques till date.

Interpreting noun compounds is a challenging task. It involves uncovering the underlying predicate which is dropped in the formation of the compound. In most cases, this predicate is of the form VERB+PREP. It has been observed that uncovering the preposition is a significant step towards uncovering the predicate. In this paper, we attempt to paraphrase noun compounds using prepositions. We consider noun compounds and their corresponding prepositional paraphrases as parallelly aligned sequences of words. This enables us to adapt different architectures from cross-lingual embedding literature. We choose the architecture where we create representations of both noun compound (source sequence) and its corresponding prepositional paraphrase (target sequence), such that their sim- ilarity is high. We use LSTMs to learn these representations. We use these representations to decide the correct preposition. Our experiments show that this approach performs considerably well on different datasets of noun compounds that are manually annotated with prepositions.

BACKGROUND: Social media is a useful platform to share health-related information due to its vast reach. This makes it a good candidate for public-health monitoring tasks, specifically for pharmacovigilance. We study the problem of extraction of Adverse-Drug-Reaction (ADR) mentions from social media, particularly from Twitter. Medical information extraction from social media is challenging, mainly due to short and highly informal nature of text, as compared to more technical and formal medical reports. METHODS: Current methods in ADR mention extraction rely on supervised learning methods, which suffer from labeled data scarcity problem. The state-of-the-art method uses deep neural networks, specifically a class of Recurrent Neural Network (RNN) which is Long-Short-Term-Memory network (LSTM). Deep neural networks, due to their large number of free parameters rely heavily on large annotated corpora for learning the end task. But in the real-world, it is hard to get large labeled data, mainly due to the heavy cost associated with the manual annotation. RESULTS: To this end, we propose a novel semi-supervised learning based RNN model, which can leverage unlabeled data also present in abundance on social media. Through experiments we demonstrate the effectiveness of our method, achieving state-of-the-art performance in ADR mention extraction. CONCLUSION: In this study, we tackle the problem of labeled data scarcity for Adverse Drug Reaction mention extraction from social media and propose a novel semi-supervised learning based method which can leverage large unlabeled corpus available in abundance on the web. Through empirical study, we demonstrate that our proposed method outperforms fully supervised learning based baseline which relies on large manually annotated corpus for a good performance.

Continuous training is crucial for creating and maintaining the right skill-profile for the industrial organization’s workforce. There is a tremendous variety in the available trainings within an organization: technical, project management, quality, leadership, domain-specific, soft-skills, etc. Hence it is important to assist the employee in choosing the best trainings, which perfectly suits her background, project needs and career goals. In this paper, we focus on algorithms for training recommendation in an industrial setting. We formalize the problem of next training recommendation, taking into account the employee’s training and work history. We present several new unsupervised sequence mining algorithms to mine the past trainings data from the organization for arriving at personalized next training recommendation. Using the real-life data about trainings of 118,587 employees over 5019 distinct trainings from a large multi-national IT organization, we show that these algorithms outperform several standard recommendation engine algorithms as well as those based on standard sequence mining algorithms.

Performance appraisal (PA) is a crucial HR process that enables an organization to periodically measure and evaluate every employee's performance and also to drive performance improvements. In this paper, we describe a novel system called HiSPEED to analyze PA data using automated statistical, datamining and text-mining techniques, to generate novel and actionable insights / patterns and to help in improving the quality and effectiveness of the PA process. The goal is to produce insights that can be used to answer (in part) the crucial "business questions" that HR executives and business leadership face in talent management. The business questions pertain to (i) improving the quality of the goal setting process, (ii) improving the quality of the self-appraisal comments and supervisor feedback comments, (iii) discovering high-quality supervisor suggestions for performance improvements (iv) discovering evidence provided by employees to support their self-assessments (v) measuring the quality of supervisor assessments (vi) understanding the root-causes of poor and exceptional performances (vii) detecting instances of personal and systemic biases and so forth. The paper discusses specially designed algorithms to answer these business questions and illustrates them by reporting the insights produced on a real-life PA dataset from a large multi-national IT services organization.

End-to-end relation extraction refers to identifying boundaries of entity mentions, entity types of these mentions and appropriate semantic relation for each pair of mentions.Traditionally, separate predictive models were trained for each of these tasks and were used in a "pipeline" fashion where output of one model is fed as input to another.But it was observed that addressing some of these tasks jointly results in better performance.We propose a single, joint neural network based model to carry out all the three tasks of boundary identification, entity type classification and relation type classification.This model is referred to as "All Word Pairs" model (AWP-NN) as it assigns an appropriate label to each word pair in a given sentence for performing end-to-end relation extraction.We also propose to refine output of the AWP-NN model by using inference in Markov Logic Networks (MLN) so that additional domain knowledge can be effectively incorporated.We demonstrate effectiveness of our approach by achieving better end-to-end relation extraction performance than all 4 previous joint modelling approaches, on the standard dataset of ACE 2004.

Towards the goal of enhancing interpretability of Latent Dirichlet Allocation (LDA) topics, we propose WikiLDA, an enhancement to LDA using Wikipedia concepts. In WikiLDA, initially, for each document in a corpus we "sprinkle" (append) its most relevant Wikipedia concepts. We then use Generalized Pólya Urn (GPU) to incorporate word-word, word-concept, and concept-concept semantic relatedness into the generative process of LDA. As the most probable concepts from inferred topics can be referred on Wikipedia, the topics are likely to become more interpretable and hence more usable in acquiring domain knowledge from humans for various text mining tasks (e.g. eliciting topic labels for text classification). Empirical results show that a projection of documents by WikiLDA in a semantically enriched and coherent topic space leads to improved performance in text classification like tasks, especially in domains where the classes are hard to separate.

The notion of roles is crucial in project management across various domains. A role indicates a broad set of tasks, activities, deliverables and responsibilities that the person needs to carry out within a project. Assigning roles to team members clarifies the expectations of work items to be delivered by each and structures the interactions of the team among themselves as well as with external stakeholders. This paper analyzes a sizeable real-life dataset regarding the actual usage of roles in software development and maintenance projects in a large multinational IT organization. The paper introduces and formalizes concepts such as seniority level of a role, career progression and career lines, formulates various business questions related to role-based project management, proposes analytics techniques to answer them and outlines the actual results produced to answer the business questions. The business questions are related to dependencies between roles, patterns in role assignments and durations, predicting role changes, discovering insights useful for meeting career aspirations, interesting role sequences etc. The proposed analytics algorithms are based on Markov models, sequence mining, classification and survival analysis.

In this paper, we propose Latent Dirichlet Allocation (LDA) [1] based document classification algorithm which does not require any labeled dataset. In our algorithm, we construct a topic model using LDA, assign one topic to one of the class labels, aggregate all the same class label topics into a single topic using the aggregation property of the Dirichlet distribution and then automatically assign a class label to each unlabeled document depending on its "closeness" to one of the aggregated topics.