Conversations with data: Advancing the state of the art in language-driven data exploration

Publié mai 3, 2021

Par Alex Polozov , Senior Researcher Chris Meek , Partner Researcher Ahmed Awadallah , Partner Research Manager

Partagez cette page

One key aspiration of AI is to develop natural and effective task-oriented conversational systems. Task-oriented conversational systems use a natural language interface to collaborate with and support people in accomplishing specific goals and activities. They go beyond chitchat conversation. For example, as personal digital assistants, they ease the stress of trip planning or reduce the expertise required to generate a sales report from a database. While natural language understanding (NLU) technology and research have achieved remarkable recent progress, task-oriented assistance requires tackling additional challenges in practical NLU.

Consider a prime application of task-oriented conversations: language-driven data exploration. Data scientists, analysts, and information workers routinely spend more than half of their time exploring, visualizing, and reformatting datasets, according to Anaconda’s “The State of Data Science 2020 (opens in new tab).” Not only time-consuming, this process is also error prone and typically requires data science programming skills, such as knowledge in Python, R, or SQL. Augmenting interactive data science environments like Microsoft Excel (opens in new tab)or Jupyter (opens in new tab) with language-driven assistance would not only save time but also democratize data exploration. For instance, an analyst could ask a system in natural language to plot last month’s sales metrics from her database rather than program a filtered visualization in SQL+R. Importantly, such a system should still allow analysts to inspect and edit program snippets after assisting with the most laborious parts of exploratory data science. This transparency and ability to edit will empower analysts and allow them to have confidence in the outcome of the work.

Data exploration highlights a core NLU challenge that plagues all task-oriented conversational systems. Understanding people’s language—and thus, their task intent—must be grounded. That is, what has been said must be interpreted relative to its context. Task-oriented systems deal with two kinds of relevant context. First, every task acts upon a structured ontology such as a database, a spreadsheet, or an API. The ontology provides data context, which influences language understanding. For example, the analyst question “Which departments have unfinished projects?” refers to “departments” and “projects” in her database and “unfinished” likely refers to a project’s status column in that database, all of which the system must emit as column references in the desired SQL program. Second, conversational systems must consider multi-turn dynamics of an interaction, which create conversation context. For example, the analyst might follow up her exploration with “What is their total budget,” implicitly referring to “unfinished projects” from the previous turn.

Publication SCoRe: Pre-Training for Context Representation in Conversational Semantic Parsing

Recently, we’ve made several fundamental contributions (opens in new tab) to these challenges. In the 2021 International Conference on Learning Representations (ICLR) (opens in new tab) publication “SCoRe: Pre-Training for Context Representation in Conversational Semantic Parsing,” (opens in new tab) we introduce SCoRe, a task-oriented conversational system with multiple applications. SCoRe achieves new state-of-the-art performance in interactive data exploration (on SParC (opens in new tab) and CoSQL (opens in new tab) benchmarks) and task-oriented dialogue (MultiWOZ (opens in new tab)), improving upon previous best techniques by up to 12 percent. SCoRe addresses the conversation context challenge through its task-oriented pretraining methodology, which learns language representations that link multiple conversation turns. To address the data context challenge, SCoRe builds upon our previous work in RAT-SQL (opens in new tab) and StruG (opens in new tab). These two publications introduce a unified framework for language understanding in the context of a structured database. It has since been leveraged in numerous applications in addition to SCoRe. We’re presenting SCoRe at ICLR on Monday, May 3, from 5 PM to 7 PM Pacific Time and 7 PM to 9 PM Pacific Time. The SCoRe code will be published on GitHub (opens in new tab); please follow the repository for updates.

Two examples of user dialogue and their corresponding formal programs and databases. The first example is a multi-turn text-to-SQL task. The user query “Find the names of the top 3 highest sales books” corresponds to the formal program “SELECT title FROM book ORDER BY sale_amount DESC LIMIT 3”. The follow-up user query, “Who are their authors,” corresponds to the formal program “SELECT t1.title, t1.name FROM author AS t1 JOIN book AS t2 ON t1.id = t2.author_id ORDER BY t2.sale_amount DESC LIMIT 3”. The next turn, “Also show the names of their publishers,” corresponds to the formal program “SELECT t1.title, t1.name, t3.name FROM author AS t1 JOIN book AS t2 ON t1.id = t2.author_id JOIN press AS t3 ON t2.press_id = t3.id ORDER BY t2.sale_amount DESC LIMIT 3”. In the corresponding database, there is an “Author” table with an “id” column, a “name” column, a “country” column, and an ellipsis signifying additional columns; a “Press” table with an “id” column, a “name” column, an “address” column, and an ellipsis signifying additional columns; and a “Book” table with an “id” column, a “title” column, an “author id” column, a “sale_amount” column, and an ellipsis signifying additional columns. The second example is a dialogue state tracking task. The user query “I am looking for a cheap restaurant in the centre of the city” corresponds to the formal program “Restaurant(Price=cheap, area=center)”. The system replies, “There is a cheap Chinese restaurant called Dojo Noodle Bar,” to which the user answers, “Yes please, for 8 people at 18:30 on Thursday,” which corresponds to the formal program “Restaurant(Price=cheap, area=center, name=Dojo Noodle Bar, people=8, time=18:30, day=Thursday)”. The user notes, “I also need to book a taxi between to the restaurant at 20:30,” which corresponds to the formal program “Restaurant(Price=cheap, area=center, name=Dojo Noodle Bar, people=8, time=18:30, day=Thursday) Taxi(leaveAt=20:30, destination=Dojo Noodle Bar)”. In the corresponding database, there is a “Restaurant” table with a “name” column, a “price” column, an “area” column, a “time” column, and an ellipsis signifying additional columns; and a “Taxi” table with a “leaveAt” column, a “destination” column, and an ellipsis signifying additional columns. — Figure 1: Achieving natural and effective task-oriented conversational systems requires the ability to interpret natural language based on data context—how a query relates to the ontology over which it is being made—and conversation context—how a query relates to previous conversational turns. Above are examples of conversational tasks from the text-to-SQL dataset SParC and the dialogue modeling dataset MultiWOZ, which exhibit different forms of such context for the system to learn.

Data context representation

The grounding challenges associated with data context and conversation context are distinct yet interconnected, and progress on both is critical to build effective task-oriented conversational systems. Here, we first address data context grounding by focusing on a single-turn version of the data exploration problem known as database question answering (DBQA). As you’ll see, techniques developed for DBQA facilitate data context grounding in broader applications of task-oriented conversational systems.

RAT-SQL: Joint representation of question and data context

When an analyst asks a question over her database, that question and its associated data context—the database schema—are both embedded into distributed intent representations by a neural encoder network. The key to addressing the data context challenge is jointly contextualizing intent representations—the question and the schema provide important context to each other. For example, in Figure 1, “Find the names of the top 3 highest sales books” refers to the title column even though the question doesn’t mention “titles.” From question language alone, author.name could superficially seem a better match.

Transformers are the most effective approach for contextualized representation learning in modern NLU. They’re based on self-attention, which, in one interpretation, learns latent relations between the inputs—in this case, question words and column/table names in the schema. While effective in multiple fields, self-attention requires large training corpora, and human-authored DBQA datasets reach only up to 10,000 training instances. With limited data, even simple natural language relations, easily discovered in machine translation and other NLU systems, can be challenging. For instance, in our DBQA experiments, standard Transformers struggled to reliably link “sales” to the column sale_amount. In the 2020 Meeting of the Association for Computational Linguistics (ACL) publication “RAT-SQL: Relation-Aware Schema Encoding and Linking for Text-to-SQL Parsers,” we introduced relation-aware Transformers (RAT), which reduce training data requirements by incorporating known relational information about the database into self-attention. They allow the encoder to consider, for instance, database foreign keys without either rediscovering them (as in standard Transformers) or hard-coding the network’s relational structure to follow them (as in graph neural networks). As such, RAT augments the learning efficacy of Transformers with rich background knowledge about relational structure.

We combine a RAT encoder with a grammar-driven program decoder into an end-to-end DBQA model called RAT-SQL. On Spider (opens in new tab), currently the most challenging text-to-SQL dataset, RAT-SQL achieved a new state of the art of 65.6 percent exact-match accuracy at the time of its publication. Relation-aware data contextualization added a more than 5 percent margin over previous best techniques. Since then, numerous researchers from other institutions have built upon RAT-SQL in their DBQA models, achieving even more impressive performance (opens in new tab).

StruG: Structure-grounded pretraining for robust language understanding

Joint representation of the question and its data context fundamentally requires solving an alignment problem—that is, linking words in the question to the data columns they reference. Relation-aware Transformers effectively incorporate known relations between the question and the data, but alignment often requires additional background knowledge either from the database content or from NLU at large. For example, in Figure 2, linking “History” to department_name is challenging without looking at the database content even though people’s real-world experience intuitively aligns these two phrases.

Two illustrations of alignment between natural language (NL) utterances and tables and column names or content of these tables. The top example shows a DBQA task. The database (DB) contains two tables: “student” and “department”. The table “student” contains columns “id”, “name”, “department_name”, “total_credits”, and an ellipsis signifying additional columns. The table “department” contains columns “id”, “name”, “building”, “budget”, and an ellipsis signifying additional columns. The NL utterance is “What is the name of the student who has the highest total credits in the History department.” The corresponding SQL program is “SELECT name FROM student WHERE department_name = ‘History’ ORDER BY total_credits DESC LIMIT 1”. The words “History department” in the NL utterance, the column name “department_name” in the DB, and the clause “department_name = ‘History’” in the SQL program are highlighted to indicate alignment. The words “total credits” in the NL utterance, the column name “total_credits” in the DB, and the clause “total_credits” in the SQL program are highlighted differently to indicate another alignment. The bottom example shows an instance of parallel text-table annotation from the web. It’s a table with the columns “train number”, “departure station”, “departure time”, “departure day”, “arrival station”, and an ellipsis signifying additional columns. It shows two rows of content. The first row has values “11417” for “train number”, “Pune Junction” for “departure station”, “22:00 PM” for “departure time”, “Thu” for “departure day”, and “Nagpur Junction” for “arrival station”. The second row has values “11418” for “train number”, “Nagpur Junction” for “departure station”, “15:00 PM” for “departure time”, “Fri” for “departure day”, and “Pune Junction” for “arrival station”. The NL utterance about the table is “The 11417 Pune-Nagpur Humsafar Express runs between Pune Junction and Nagpur Junction.” The word “11417” in the NL utterance, the column name “train number”, and its first-row value “11417” in the table are highlighted to indicate alignment. The words “Pune Junction” in the NL utterance, the column name “departure station”, and its first-row value “Pune Junction” in the table are highlighted differently to indicate a second alignment. The words “Nagpur Junction” in the NL utterance, the column name “arrival station”, and its first-row value “Nagpur Junction” in the table are highlighted differently to indicate a third alignment. — Figure 2: For a task-oriented conversational system to account for data context, it must be able to link words in a query to their corresponding data columns based on background knowledge about the relations between the question and the data. The relations align words in a query to column names or table content, as illustrated here in database question answering (top) and in parallel text-table corpora from the web (bottom).

Background knowledge on question-table alignment naturally occurs in parallel text-table corpora such as ToTTo. They pair data tables with relevant utterances about them, such as table summaries and data references. While such utterances are not typically in question form, they exhibit the same alignment patterns as questions in DBQA. In an upcoming 2021 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL) publication “Structure-Grounded Pretraining for Text-to-SQL,” we propose a method to leverage this data for pretraining contextual representation models. StruG, short for “Structure-Grounded pretraining,” introduces three critical pretraining tasks that use text-table alignment annotations as weak supervision:

Column grounding: given a column—for example, train number in Figure 2—predict whether it’s relevant to the utterance
Value grounding: given a word in the utterance—for example, “11417”—predict whether it refers to a cell in some column
Column-value mapping: given a word—such as “11417”—and a column name—train number—predict whether they align

Importantly, the StruG model can’t observe database content when predicting these tasks. It learns contextualized alignment solely from the utterance and the data schema. After pretraining, we apply StruG in DBQA to emit initial question/table representations that communicate information about the context structure to the downstream RAT-SQL model.

On the Spider dataset, StruG-augmented RAT-SQL performs competitively with all state-of-the-art models even without using database content. More importantly, its learned text-table alignment makes database question answering more robust. When development-set questions in Spider are rephrased in a more realistic, fluid natural language, execution accuracy of state-of-the-art models drops by 11 to 20 percent, but StruG-augmented RAT-SQL never suffers more than a 10 percent loss.

SCoRe: Pretraining for conversation context representation

The solutions to the data contextualization challenge are also relevant for the conversation contextualization challenge. To account for multi-turn dynamics of dialogue, a conversational system must ground an individual’s question in both its data context—that is, its ontology—and in the questions from previous turns. SCoRe introduces a task-oriented pretraining methodology to encode both.

SCoRe pretrains a task-oriented language model contextualized by the conversational flow and the underlying ontology. In pretraining, the SCoRe model is self-supervised by two novel task-oriented objectives in addition to the established masked language model (MLM) objective (opens in new tab). These objectives facilitate the accurate representation of the conversational flow between dialogue turns and how this flow maps to the desired columns in the ontology. For example, in the first question of Figure 1, “Find the names of the top 3 highest sales books,” the model needs to apply the order by operation to the column sale_amount to find the books with the highest sales. In the follow-up question “Who are their authors,” the model needs to understand that it should maintain the context of the previous question while also selecting a new column, name, from the author table.

The first pretraining objective of SCoRe, Column Contextual Semantics (CCS), aligns the question with the ontology. For each column in the ontology, CCS trains the model to predict the operations that should be performed on this column in each conversational turn. Specifically, SCoRe uses the encoding of each column or table name to predict its corresponding operation. The second pretraining objective, Turn Contextual Switch (TCS), captures the conversational flow and how it’s grounded in the model’s expected output programs. It aims to predict the difference in programs from different dialogue turns based on the corresponding user questions. For example, the current turn may differ from the previous one by adding an additional filtering condition or changing the order of the results.

A diagram of the pretraining of a SCoRe encoder. It shows a “Transformer Encoder” block that takes three blocks of input—“Current Question,” “Dialogue History,” and “Database Schema”—and produces multiple outputs. The “Current Question” block contains the sentence “also show the names of their publishers” preceded by a special separator token. The “Dialogue History” block contains the two sentences “who are … authors” and “find <mask> … books” delimited by separator tokens with ellipsis signifying additional words in the middle and “<mask>” signifying a special masked-word token. The “Database Schema” block contains the column names “author id”, “author name”, and “sale amount”, each surrounded by separator tokens, and an ellipsis signifying additional columns. The first two outputs are for the TCS objective. They’re aligned with the separator token from the current question and from the first sentence of the dialogue history, respectively, and show the target output “INS(SELECT.column)” and “INS(SELECT.column)”, respectively. The third output is for the MLM objective. It’s aligned with the masked-word token in the dialogue history and shows the target output “the”, which is the word that was masked. The final three outputs are for the CCS objective with an ellipsis signifying additional CCS outputs. They’re aligned with column names “author id”, “author name”, and “sale amount”, respectively, in the database schema. They show the target outputs “None”, “SELECT”, and “ORDER BY DESC LIMIT”, respectively. — Figure 3: The SCoRe model is a Transformer-based encoder that takes as input the current natural language question along with its two kinds of context: *conversation context,* or its dialogue history of previous questions, and *data context,* or its database schema or other ontology. To encode this joint input into informative distributed representations, SCoRe is trained to predict three objectives. For each dialogue question, the Turn Contextual Switch (TCS) objective predicts the expected difference between the desired formal program and the program for the previous turn’s question. For each column in the data context, the Column Contextual Semantics (CCS) objective predicts a desired operation that should be performed upon this column in the desired formal program, if any. Finally, the standard MLM objective reconstructs words in the input that were masked with a special token.

State-of-the-art results

Our empirical results show that SCoRe can be effectively used as a feature representation encoder with strong baseline models for a wide range of tasks and can significantly improve the performance of existing strong baseline models by simply replacing an existing pretrained language model with our SCoRe pretrained model.

SCoRe achieves state-of-the-art results when evaluated on three popular benchmarks for task-oriented conversational systems: SParC (opens in new tab) (sequential text-to-SQL), CoSQL (opens in new tab) (conversational text-to-SQL), and MultiWOZ (opens in new tab) (dialogue state tracking). It also performs competitively with state-of-the-art techniques on SQA (opens in new tab) (sequential question answering). Moreover, SCoRe delivers even larger improvements when in-domain data is limited—for example, in a low-resource setting where only 10 percent of the training data is available. This wide range of applications demonstrates the effectiveness of addressing both context grounding challenges jointly.

Revolutionizing interaction with data

Task-oriented conversational systems can revolutionize people’s natural interaction with structured data and APIs. Using natural language as a universal interface has been a major goal of human-computer interaction and knowledge management fields for decades. Early attempts have faced challenges because of limitations in language understanding capability, extensibility, and transparency, among other areas. However, recent years have seen a major resurgence powered by interest in impactful applications such as personal digital assistants, question answering systems, automatic reporting, and AI-assisted data science.

Many challenges remain. Conversational interfaces also require us to make systems reasoning and results explainable and trustworthy to those using them. Language-driven exploration must be supported by interactive interfaces for debugging and correcting generated programs or the underlying dataset. Creating systems that interact with those using them to resolve knowledge gaps and continue to learn to reduce human intervention over time remains an open research challenge. We’ve studied additional forms of interaction to address these challenges in language-driven data exploration, including natural language feedback for correcting misinterpretations and Debug-It-Yourself (DIY) multimodal feedback for assessing a system’s responses and fixing errors in an interactive user interface. Both improve systems’ accuracy and transparency, yet more research is needed to integrate them into task-oriented conversational systems more broadly. We hope that SPLASH, a dataset of utterances, misinterpretations, and corrections we created from our feedback studies, will prove useful to facilitate that research.

Finally, many NLU challenges stem from the limitations of current benchmarks and datasets of task-oriented conversational systems. As we scale our techniques and integrate them into real-world applications, we’ll encounter more realistic scenarios and workflows, which will undoubtedly expose new research challenges. As such, language-driven data exploration will not only be one of the most impactful applications for the field, but also the catalyst to its further progress.

Acknowledgment

The development of SCoRe is a result of the collaborative efforts of Tao Yu (opens in new tab) of Yale University, Rui Zhang (opens in new tab) of The Pennsylvania State University, Microsoft researchers Alex Polozov (opens in new tab) and Chris Meek (opens in new tab), and Microsoft Senior Principal Research Manager Ahmed H. Awadallah (opens in new tab). RAT-SQL was led by Bailin Wang (opens in new tab) of The University of Edinburgh in collaboration with Microsoft researchers Richard Shin (opens in new tab), Xiaodong Liu (opens in new tab), and Polozov and former Microsoft researcher Matthew Richardson (opens in new tab). StruG was led by Xiang Deng (opens in new tab) of The Ohio State University in collaboration with Awadallah, Meek, Polozov, Richardson, and assistant professor Huan Sun (opens in new tab). Yu, Wang, and Deng conducted the work during their Microsoft Research internships.