[Paper Review] Exploring Conversational Language Generation for Rich Content about Hotels

Marilyn Walker, Albry Smither+3 more|arXiv (Cornell University)|May 1, 2018

Topic Modeling10 references4 citations

TL;DR

This paper proposes a framework for generating natural, conversational language about hotels using rich, structured content typically found in hotel descriptions. By collecting and analyzing three types of dialogue data—self-dialogues, meaning representation-based generation, and crowd-sourced dialogues—it identifies key stylistic differences between formal written descriptions and conversational speech, enabling improved natural language generation and retrieval for hotel dialogue systems.

ABSTRACT

Dialogue systems for hotel and tourist information have typically simplified the richness of the domain, focusing system utterances on only a few selected attributes such as price, location and type of rooms. However, much more content is typically available for hotels, often as many as 50 distinct instantiated attributes for an individual entity. New methods are needed to use this content to generate natural dialogues for hotel information, and in general for any domain with such rich complex content. We describe three experiments aimed at collecting data that can inform an NLG for hotels dialogues, and show, not surprisingly, that the sentences in the original written hotel descriptions provided on webpages for each hotel are stylistically not a very good match for conversational interaction. We quantify the stylistic features that characterize the differences between the original textual data and the collected dialogic data. We plan to use these in stylistic models for generation, and for scoring retrieved utterances for use in hotel dialogues

Motivation & Objective

• To address the gap in conversational agents' ability to discuss rich hotel content beyond basic attributes like price and location.
• To investigate how written hotel descriptions differ stylistically from natural conversational language in dialogue systems.
• To collect and analyze diverse dialogue data to inform the design of natural language generation models for complex hotel information.
• To develop stylistic models and scoring functions for selecting appropriate utterances in multi-turn hotel dialogues.
• To provide a publicly available corpus for training and evaluating conversational agents in the hotel domain.

Proposed method

• Conducted three data collection experiments: self-dialogue generation, meaning representation-based utterance generation, and crowd-sourced multi-turn dialogues.
• Used linguistic analysis tools (LIWC) to quantify stylistic differences across datasets, focusing on features like personal pronouns, social processes, and focus present.
• Compared original InfoBox descriptions (ORIG) with generated (PARA), combined (PROP+ROOM), and dialogic (DIAL) utterances to identify stylistic divergence.
• Identified key stylistic features—such as higher use of personal pronouns and social processes in dialogic data—that distinguish conversational from descriptive language.
• Proposed training a 'conversational style ranker' using identified linguistic features to retrieve contextually appropriate utterances.
• Planned to templatize collected utterances and index them for context-aware retrieval in dialogue systems.

Experimental results

Research questions

• RQ1 How do the linguistic styles of written hotel descriptions differ from those of natural conversational utterances in hotel dialogues?
• RQ2 Which data collection method produces the most natural and contextually appropriate utterances for hotel dialogue systems?
• RQ3 What specific stylistic features distinguish conversational language from formal written descriptions in the hotel domain?
• RQ4 Can stylistic features be used to train a model that ranks or retrieves high-quality, contextually relevant utterances for dialogue systems?
• RQ5 How can rich, structured hotel content be effectively transformed into fluent, natural, and context-sensitive dialogue turns?

Key findings

• The original written hotel descriptions (ORIG) exhibit significantly lower use of personal pronouns (0.06) and social processes (4.81) compared to dialogic data (DIAL), which show 10.17 and 14.66 respectively.
• The DIAL condition demonstrated the highest level of conversational features, including focus present (14.4) and social processes (14.66), indicating stronger alignment with natural dialogue.
• The PARA and PROP+ROOM conditions showed intermediate stylistic profiles, with higher personal pronoun use than ORIG but lower than DIAL, suggesting partial alignment with conversation.
• The self-dialogue collection produced the most natural-sounding utterances but required the highest cost and posed challenges for direct reuse.
• The generation from meaning representations produced high-quality utterances but lacked contextual sensitivity, indicating a need for context-aware refinement.
• The corpus of collected data—spanning multiple styles and conditions—is publicly available at nlds.soe.ucsc.edu/hotels for research and model training.