Human Language Technologies

Voice of Customer Analytics
IBM Voice of Customer Analytics (VoCA) is a managed service in the area of Customer Relationship Management (CRM) analytics that provides actionable business insights for customers and operational efficiency. In VOCA, heterogeneous structured and unstructured data sources like customer profile and transactions, customer satisfaction surveys, call transcripts, agent logs, and activity records are captured and linked together. An IBM Business Analyst analyzes various heterogenous data sources, enabled by advanced analytical capabilities such as data linking, text clustering, text annotation, sentiment mining and predictive modeling, to come up with actionable insights regarding customer churn, first call resolution, key customer satisfaction and dissatisfaction drivers.

Researchers: Shantanu Godbole, Ajay Gupta, Himabindu Lakkaraju, Virendra Varshneya, Ashish Verma

Prospect
Hiring and staffing of skilled resources is a crucial activity in large services organizations. The hiring function faces the tremendous challenge of matching hundreds of applications against various job openings to arrive at a short-list of skilled candidates for interview. With limited bandwidth in terms of manpower and increasing number of candidates, recruiters are faced with conflicting objectives of quality and timeliness in the hiring process. The goal of this project at IBM Research - India is to develop a decision support tool leveraging advance text analytics, to help recruiters identify the most suitable candidates easily. Some of the challenges involved in such tasks include error-free extraction of candidate information(skill, education detail, past experience) from resumes (written in non-standard formats), inferring skill level of a person from 'skill mentions', matching and choosing best skill levels for various job requirements, etc. The researchers work closely with the HR organization of IBM to develop this technology.

Researchers: Amit K Singh, Rose C Kanjirathinkal, Karthik Visweswariah

Sensei
Sensei is a web-enabled tool that evaluates a person's spoken English skills. It uses advanced speech processing techniques to evaluate various parameters, including pronunciation, grammar, syllable stress and spoken fluency. Sensei enables people to improve their English speaking skills via the Internet and provides a score that can be used to determine their skill levels in spoken English. As a highly scalable and cost effective solution, Sensei can help organizations, especially those in customer service, improve the language skills of their employees. It is also valuable to organizations in the language learning domain.

Researchers: Ashish Verma, Om Deshmukh, Harish Doddala

Workforce Optimization 
In large IT services organizations, effective matching of practitioners to project openings is critical to profitability. Practitioners and project openings are characterized by structured attributes such as job role, skill set, location city and unstructured attributes such as job description and resumes. Complexity arises due to inter-relationships between attributes (for example the hierarchical relationship between skills), difficulty in extracting the level of experience from resumes, missing or incorrect data, as well as aligning the process to customer requirements. Workforce Optimization is an interdisciplinary project that uses constrained optimization, machine learning, and text analytics techniques to develop a decision support solution which not only provides effective matching of practitioners to projects but also reduces overhead involved in information sharing by end users.

Researchers: Gyana Parija, Pranav Gupta, Rohit Lotlikar, Nanda Kambhatla

Effective knowledge management for services
Improving productivity of practitioners is a critical goal for IT services organizations. In large organizations with many teams providing services to different clients, it is important to effectively re-use the knowledge acquired by practitioners.. The IBM team is working on developing a technology for effective knowledge management that will enable practitioners to re-use the knowledge in their day-to-day activities. It will also provide a general framework for analyzing problem data to automatically identify the main challenge areas for the concerned client. This work involves some interesting and challenging problems in the areas of information retrieval, information extraction, machine learning and data management. The IBM Research team works very closely with the technical team in IBM's services unit as well as the services teams that directly support clients.

Researchers: Debapriyo Majumdar, Rose C Kanjirathinkal, Virendra Varshneya, Karthik Visweswariah

TRAIL - Machine translation for Indian Languages
We are building Statistical Machine Translation Systems for translating documents between selected Indian Languages and English. Initial language pairs of interest include Hindi/English and Urdu/English. The core translation technology enables various applications such as cross lingual search, translation of chat messages and serves as an efficiency enhancer for human translators.

Researchers: Karthik Visweswariah, Nanda Kambhatla, Anand Ramanathan, Ankur Gandhe, Rashmi Gangadharaiah

MNBA – Multi-channel Next Best Action for Customer Relationship Management
We are building a system for integrating different channels of interaction through which customers interact with a business organization. The system would provide a seamless real-time presentation of all interactions for a customer, analyze the interactions using voice and text analysis techniques to extract important information, and devise a strategy for personalized next best action from the organizations point of view, for enhanced customer relationship management. This work involves some very challenging and current problems in the areas of information extraction, sentiment mining and predictive modelling of customer interests based on past transactions.

Researchers: Rohit Lotlikar, Nanda Kambhatla, Angshu Rai, Sneha Chaudhari

Whole Call Analytics
The objective of this project is to build technology and solutions that help in understanding transcriptions of calls made to call centers by customers. We are working on methods to model conversations between customers and agents with an aim to discover different topics, relations among topics, and flow of topics in an unsupervised manner. We are also exploring semi-supervised techniques that can be used to incorporate prior knowledge in these methods. The ASR output of transcriptions is very noisy and therefore we are also working on methods for clustering and classification with noisy and inaccurate data.

Researchers: Sachindra Joshi, Shajith Iqbal, Ashish Verma

SCION
Social media has provided a very efficient platform to people who want to express their opinions and discuss issues regarding organizations/people/places/governments and other entities that they interact with. This project aims at mining these community conversations and interactions on social media websites. The goal of the project is to develop techniques for automatically discovering important concepts/topics that people talk about, influential authors, determining important posts/blogs after filtering out spam and categorizing these posts into different topics. These techniques are useful for a variety of applications that involve analyzing social media conversations.

Researchers: Jitendra Ajmera, Himabindu Lakkaraju, Nanda Kambhatla (IRL), Meenakshi Nagarajan, Hyung-il Ahn, Matthew Denesuk (ARC)

Real Time Transcription & Agent Assistance (RTTŠAA)
The goal of this project is to develop a real-time speech analytics solution for assisting and monitoring the agent in during a customer service interaction. Some of the research challenges involve achieving high recognition accuracy on highly noisy call center audio data and applying text analytics in real-time to predict potential problematic calls and generate call summaries. RTTAA also attempts to unlock the business intelligence from the call data by identifying the problems/issues to automatically suggest relevant solutions from a solution/knowledge repository thereby reducing the average handling time (AHT) and leading to a better customer experience.

Researchers: Harish Doddala, Ashish Verma

iSAGE
In current age, users have various means of finding solutions to problems they face, pertaining to products they use. It ranges from modern age search engines at one end to complex human powered contact centers at another. One aspect that is common across these systems is for most of the problems, finding the resolution is an interactive process. because questions posed by users are often underspecified thus having a broad space of possible answers. Therefore, clarification dialogue is often needed to negotiate with the user the exact scope and intent of the problem.
While humans are generally good at it, interactive process is a great challenge for automated systems. The goal of the FRR project iSAGE, is to build an interactive system for problem solving with minimal knowledge engineering effort required for adapting it to a new domain. As distinguished from existing system, goal is to return solution to the problem rather than a list of webpage links.
iSAGE is an interactive problem resolution system that allows users to pose complex questions in near-natural language. It targets questions which need not be factual, but questions for which solution may include procedure or a series of steps to be executed. For doing this it mines information from various sources (web forums, technical documents, product manuals, chat transcripts) and engages in a interactive dialogue to solve the users problem

Researchers: Amit K Singh, Rose C Kanjirathinkal, Rashmi Gangadharaiah, Dinesh Raghu, Srujana Merugu, Sachindra Joshi, Nanda Kambhatla, Karthik Visweswariah, Ashish Verma

Global Delivery Assist
While global delivery of services is becoming prevalent, language is still a bottleneck in dictating the location from where services can be delivered. We are working on a tool that would provide high quality domain specific translation to aid in translating documents that aids in the delivery of software services. The tool continually learns from past translations and adapts to the domain in question while also storing past translations in a translation memory to allow collaboration between translators for greater efficiency in the translation process.

Researchers: Virendra Varshneya, Karthik Visweswariah, Nanda Kambhatla, Deepti Purandare, Asha Andrews, David Lima

Recruitment Analytics
In this initiative we are analyzing various sub-steps & processes associated with the recruitment cycle to identify areas of improvement and developing Machine Learning and Optimization Methods to handle the problems. In particular we are looking to improve the following:

• Candidate Sourcing while minimizing the cost.
• Increasing candidate yield while reducing interview cost
• Identifying low attrition risk candidates
• Scheduling Interviews
• Matching Candidates to technical panels.

The overall project aim to reduce the cost of recruitment while increasing yield by identifying candidates with high onboard probability. Moreover, by identifying high attrition risk candidate, the employee turnover, re-hire and training cost can also be minimized. We have developed a tool which jointly ranks the candidate based on technical match, onboard probability and early attrition risk. The joint ranking helps to identify most desirable candidates and reduce the effort to find them through expensive process of interviews.

Researchers: Mehta, Amit K Singh, Rakesh R Pimplikar, Karthik Visweswariah