Source-linked AI summary
A Knowledge-Grounded Neural Conversation Model
Marjan Ghazvininejad, Chris Brockett, Ming-Wei Chang, Bill Dolan, Jianfeng Gao, Wen-tau Yih, Michel Galley
TL;DR
Existing data-driven conversation models often lack access to the factual and entity-grounded information needed for substantive responses. This paper introduces a knowledge-grounded extension of SEQ2SEQ that conditions generation on conversation history and retrieved textual facts, improving informativeness in human evaluation over a competitive baseline.
Problem
Fully data-driven conversation models lack grounding in external knowledge, while conversational datasets do not represent much of the world’s entity-specific information.
Method
The model generalizes SEQ2SEQ by identifying conversational entities, retrieving relevant textual facts, and combining conversational and non-conversational data through multi-task learning.
Results
Human evaluation found significant improvements in informativeness over a competitive large-scale SEQ2SEQ model baseline.
Takeaways & Limitations
The approach supports scalable, fully data-driven, open-domain conversation that exploits external knowledge without explicit slot filling.
Takeaways & Limitations
The model’s “facts” are authored text snippets that may contain subjective or inaccurate information, and exact-match grounding is high precision but low recall.
Abstract
from arXiv · showhide
Neural network models are capable of generating extremely natural sounding conversational interactions. Nevertheless, these models have yet to demonstrate that they can incorporate content in the form of factual information or entity-grounded opinion that would enable them to serve in more task-oriented conversational applications. This paper presents a novel, fully data-driven, and knowledge-grounded neural conversation model aimed at producing more contentful responses without slot filling. We generalize the widely-used Seq2Seq approach by conditioning responses on both conversation history and external "facts", allowing the model to be versatile and applicable in an open-domain setting. Our approach yields significant improvements over a competitive Seq2Seq baseline. Human judges found that our outputs are significantly more informative.
Introduction
Fully data-driven conversation models can respond appropriately but often lack factual, entity-specific content. This paper addresses that gap by combining conversational history with external textual facts to generate more informative responses.
- Introduction: Fully data-driven systems often produce conversationally appropriate responses without substantive factual content.They lack access to external knowledge, making entity-focused responses difficult.
- Introduction: Human responses can add entity-specific recommendations that neural models commonly omit.The example contrasts “Have a great time!” with a recommendation for omasake at Kusakabe.
- Introduction: The model conditions responses on both conversation history and relevant external facts, generalizing the SEQ2SEQ approach.It combines conversational and non-conversational data through techniques such as multi-task learning.
- Introduction: Conversational corpora sparsely represent named entities, whereas non-conversational texts represent them more exhaustively.Figure 2 motivates using sources such as reviews and other external text to supply contentful words.
- Introduction: 23M Twitter conversations and 1.1M Foursquare tips supported systems that significantly improved informativeness over a competitive large-scale SEQ2SEQ baseline in human evaluation.The approach is designed for versatility and open-domain applicability.
Related Work
Prior neural dialog work includes data-driven conversation generation, goal-directed slot-filling systems, and memory-network approaches. This paper differs by pursuing informative conversation without explicit slot filling and by leveraging much larger free-form text resources.
- Related Work: Data-driven conversation generation extends statistical and neural translation models from conversational data.Contextual models later incorporated conversational history, and this work extends that direction with textual side information.
- Related Work: Goal-directed neural dialog systems explicitly learn question-answer slots from relatively small task-specific datasets.The paper instead explores less clearly goal-directed but informational conversation without explicit slot filling.
- Related Work: Memory-network approaches have been used for restaurant reservations and dialog-state management with structured knowledge bases.These approaches use keywords or memory mechanisms to manage entities or dialog state.
- Related Work: Free-form text datasets can be several orders of magnitude larger than the relatively small datasets used in related memory-network work.The larger resources provide greater domain and form diversity for learning a more robust conversational backbone.
Grounded Response Generation
The model addresses the limited world knowledge of data-driven conversation systems by retrieving external text facts and conditioning responses on both facts and conversation history. Its architecture uses separate encoders and an RNN decoder, trained with multi-task learning to generate informative responses without explicit slot filling.
- Existing conversational datasets omit much world knowledge, limiting data-driven models’ ability to respond substantively about entities.
- The system retrieves text facts relevant to a conversation’s entity focus and feeds them alongside conversation history into distinct neural encoders.Facts are indexed by named entities; the focus can be found by keyword matching, entity linking, or named entity recognition.
- The approach generalizes beyond entity-specific conversational training because retrieved facts can support responses for out-of-vocabulary entities and new information.This supports a more general response-generation framework than SEQ2SEQ for datasets linking named entities to free-form text.
- Multi-task learning combines fact-grounded examples with conversation-only examples, allowing separate training conditions and conversational pretraining.The conversation-only task uses (S, R) examples, while the fact-grounded task uses ({f1, ..., fk, S}, R) examples.
- The facts encoder adapts memory networks by weighting entity-relevant facts using an RNN representation of the conversation history.Unlike original memory networks’ bag-of-words input, the RNN encoder captures interlexical dependencies.
- The decoder initializes its hidden state from a combined representation of the conversation and external facts, then generates the response word by word.The combined representation is formed by summing the dialog and fact encodings; summation performed best among tested operations, and a one-layer memory network was used.
Datasets
The experiments combine large-scale general Twitter conversations with Foursquare business tips and grounded Twitter conversations linked to Foursquare entities. Evaluation data were filtered to emphasize responses that are informative and useful.
- The study uses Twitter for conversational data and Foursquare for non-conversational entity-associated text.The approach is presented as applicable to datasets mapping named entities to free-form text, including sources such as Wikipedia, IMDB, and TripAdvisor.
- The Foursquare corpus contains 1.1M tips about restaurants and other North American establishments, including descriptions and recommendations.
- The general Twitter dataset contains 23M three-turn conversations used to learn conversational structure without associated facts.
- A separate grounded dataset contains approximately 1M two-turn Twitter conversations whose entities link to Foursquare.Links were identified through business handles or matching hashtags, with some user-generated responses removed.
- For each grounded conversation, the system retains the 10 Foursquare tips most similar to the input after retrieving potentially as many as 10k tips per handle.Similarity is computed using cosine similarity between tf-idf-weighted word-count vectors.
- The dev and test sets contain 4k conversations selected to emphasize informative and useful responses rather than purely socializing replies.
- Human-evaluation data were sampled from conversations ranked by language-model and χ-square content scores, plus random samples.The procedure selected 15k conversations under each ranking method, 15k randomly, then sampled 10k for judging.
Experimental Setup
The experimental setup compares SEQ2SEQ and multi-task variants that expose models to conversation-only, fact-grounded, and fact-reconstruction tasks. Models use large recurrent architectures, beam-search decoding, and automatic metrics including perplexity, BLEU, and lexical diversity.
- Multi-Task Learning: The training tasks are FACTS, NOFACTS, and AUTOENCODER, respectively pairing conversation and facts, conversation alone, or facts as reconstruction targets.AUTOENCODER training uses ({f1, ..., fn, S}, fi) examples and has n times as many samples as FACTS.
- Multi-Task Learning: AUTOENCODER training was found to inject more factual content into responses, while FACTS and NOFACTS represent intended operation.
- Multi-Task Learning: The SEQ2SEQ baseline uses NOFACTS training on the 23M general conversation dataset, whereas MTASK isolates multi-task effects without facts.MTASK uses NOFACTS on both the general and grounded datasets; MTASK-R adds FACTS on the grounded dataset.
- Multi-Task Learning: MTASK-F combines general NOFACTS training with grounded AUTOENCODER training, while MTASK-RF adds grounded FACTS training to those tasks.
- Implementation: The models use one memory-network layer with two-layer GRU encoders and decoders, 512 hidden cells per layer, 512-dimensional embeddings, and 1024-dimensional memory representations.Training uses Adam with learning rate 0.1, batch size 128, and gradient clipping at 5.
- Multi-Task Learning: Task batches are sampled according to mixing ratios, with all examples in a batch drawn from one task.
- Decoding: Beam search uses beam size 200 and maximum response length 30, ranking candidates with decoder likelihood, word count, and source-given-response likelihood.The reverse likelihood feature is intended to reduce generic responses such as “I don’t know”; its parameters are tuned with MERT by optimizing BLEU.
- Evaluation: Evaluation reports corpus-level BLEU, perplexity, and lexical diversity as automatic measures, with corpus-level BLEU chosen for better correlation with human judgments.
Results
The evaluation combines automatic metrics with crowdsourced judgments of appropriateness and informativeness. MTASK-R improves BLEU and lexical diversity over SEQ2SEQ, while human evaluation finds a significant informativeness gain and preserves contextual appropriateness.
- Automatic Evaluation: Injecting more factual content in MTASK-F and MTASK-RF increases perplexity, especially on grounded data.
- Automatic Evaluation: MTASK-R increases BLEU by 96% and 1-gram diversity by 71% compared with the competitive SEQ2SEQ baseline.
- Human Evaluation: 500 paired conversations were evaluated by seven judges on appropriateness and informativeness using crowdsourced comparisons.
- Human Evaluation: MTASK-R significantly outperforms SEQ2SEQ on informativeness (p = 0.003), while its appropriateness gain is small and not statistically significant.
- Human Evaluation: MTASK-RF is slightly better than SEQ2SEQ on informativeness but worse on appropriateness, with neither difference statistically significant at α = 0.05.
- Human Evaluation: The judge-preference distributions for informativeness are sharper and more distinctive than those for appropriateness.
Discussion
Examples show that grounded responses can combine factual information with contextual appropriateness, although irrelevant or contradictory facts produce errors. The main system is generally on par with SEQ2SEQ in appropriateness while improving informativeness.
- Qualitative Analysis: MTASK-RF examples generally preserve contextual appropriateness while incorporating information from external facts.
- Human Evaluation: Table 3 reports mean judgment differences with 95% confidence intervals, and boldface marks comparisons significantly better than their comparator.
- Qualitative Analysis: Fact-influenced responses may contain one unedited fact or combine text from multiple facts.
- Qualitative Analysis: The two main reported error causes are irrelevant retrieved facts and combinations of self-contradictory facts.
- Discussion: Judges found the best grounded system generally on par with SEQ2SEQ in appropriateness while significantly improving informativeness.
Conclusions
The paper presents a scalable, fully data-driven neural conversation model that exploits external knowledge without explicit slot filling. Its entity-matching grounding supports informative, versatile, open-domain conversational systems.
- The model exploits external knowledge without explicit slot filling in a scalable, fully data-driven neural conversation framework.
- Multi-task learning combines conversational and non-conversational data while generalizing the SEQ2SEQ approach.
- Simple entity matching grounds external information in conversation context, supporting informative and versatile open-domain systems.
- The engine could serve as a core component of multi-turn recommendation or conversational question-answering systems.