End to End Dialog System for Telugu

Prathyusha Danda, Prathyusha Jwalapuram, Manish ShrivastavaICON 2017 · OpenReview

Abstract

This paper presents a hybrid end-to-end dialog system for Telugu, a language often sidelined in NLP due to a lack of large-scale datasets. By combining Sequence-to-Sequence (Seq2Seq) models for natural language generation with Memory Networks for structured API querying, the system successfully navigates tourist information requests with near-perfect accuracy on synthetic benchmarks.

Introduction

Building task-oriented dialog systems—such as virtual assistants or customer service bots—typically requires massive amounts of human-annotated conversation data. For "low-resource" languages like Telugu, this data is largely non-existent. Previous attempts at Telugu dialog systems were primarily rule-based, making them brittle, difficult to scale, and unable to handle the nuances of natural conversation.

Researchers at IIIT Hyderabad addressed two main hurdles in this work:

  1. Data Scarcity: Developing a methodology to train data-hungry neural networks in the absence of a pre-existing corpus.
  2. Functional Precision: Ensuring a model can maintain a natural "chat" flow while simultaneously performing precise technical tasks, such as generating structured API calls to a database without syntax errors.

Methods

The authors propose a hybrid pipeline that effectively separates conversational interaction from functional execution. The system architecture relies on two distinct neural components working in tandem.

Hybrid system architecture combining Seq2Seq and Memory Networks.

Sequence-to-Sequence (Seq2Seq) Model

The system uses an Encoder-Decoder architecture with Gated Recurrent Units (GRU) and an attention mechanism. This component is responsible for generating natural language responses. For instance, if a user provides an area but not a place type, the Seq2Seq model generates the Telugu equivalent of "Which type of place are you interested in?"

Memory Networks

While Seq2Seq models excel at fluid text generation, they often struggle with the rigid, precise structure required for API calls (e.g., api_call banjara_hills zoo 15). To solve this, the authors utilize a Memory Network to treat API generation as a classification task rather than a generation task. The network "remembers" the dialog history and selects the most appropriate API call from a predefined set of candidates.

Deterministic Slot Filling

The system is designed to fill three specific "slots"—Area, Type of place, and Tour duration—before it attempts to query the knowledge base. This ensures that the API call has all the necessary parameters to return a valid result.

Datasets and Experimental Setup

Because no natural Telugu dialog corpus was available, the authors created a comprehensive synthetic dataset. They first built a Knowledge Base (KB) of tourist locations in Hyderabad, including attributes like area, type (historical, religious, etc.), rating, and contact information.

Using 68 user intent patterns and 9 bot response patterns, they automatically generated thousands of dialogs to simulate real-world interactions. The final training set was substantial, including 180,000 instances for the Seq2Seq model and 24,000 instances for the Memory Network.

Key Insights and Ideas

  • The API Generation Challenge: The researchers found that standard Seq2Seq models performed poorly (around 85% accuracy) when generating API calls because place names vary significantly in length and complexity. By switching to Memory Networks—which pick a response rather than generating it word-by-word—they achieved 100% accuracy for API retrieval.
  • Cross-Lingual Bridge: While the conversation happens in Telugu, the underlying API calls and Knowledge Base are maintained in English. This demonstrates a viable strategy for building local-language interfaces on top of existing English-centric data infrastructure.
  • The Importance of Context: The study quantitatively proved that adding dialog history (context) significantly boosts performance. Accuracy for system responses jumped from approximately 60% with no context to 85.5% when context was included in the Seq2Seq model.

Key Results

The hybrid system demonstrated remarkable performance on the synthetic test sets. When the task was limited to generating the next conversational sentence (excluding the API call), the Seq2Seq model achieved 100% accuracy.

Sample multi-turn dialog in Telugu leading to an API call.

The Memory Network achieved 100% accuracy for initial information retrieval and providing additional details (like phone numbers), and 99.93% for updating existing queries. As shown in the sample dialog above, the system successfully manages multi-turn interactions, prompting users for missing information until it has enough data to execute a search.

Limitations and Future Work

Despite the high accuracy, the authors note the "Template Trap." The system was trained and tested on synthetic, template-based data. Real-world human speech is much more varied, often ungrammatical, and unpredictable. Accuracy is expected to decrease when the system is exposed to "wild" user input.

Currently, the system is restricted to the tourism domain in Hyderabad. Future directions include incorporating more proactive features—such as the bot offering alternative suggestions if a query yields no results—and testing the system with native Telugu speakers to gather a more natural, non-synthetic corpus.

Summary updated

BibTeX
@inproceedings{danda2017end,
  author = {Prathyusha Danda and Prathyusha Jwalapuram and Manish Shrivastava},
  title = {{End to End Dialog System for Telugu}},
  booktitle = {ICON},
  year = {2017},
  url = {https://openreview.net/forum?id=k2CzoXP1wr},
}
Figure