Kelsey Rook
Dialogue and written documents are two prominent methods of expressing underlying ideas and exchanging information. Whereas documents take more consideration to produce and are more polished and structurally consistent, human dialogue is spontaneous, less considered, and inherently messy. Despite this inconsistency, conversation remains a rich source of information. From casual chit-chat to task-oriented interactions and formal meetings, dialogue represents a massive source of knowledge exchange.
Information Extraction (IE) is a critical sub-discipline within the field of Natural Language Processing (NLP) that focuses on converting vast quantities of unstructured text into structured, machine-readable knowledge. This process of structuring information serves as a foundational pillar for a wide array of downstream applications, including the construction of knowledge graphs (KGs). In the context of dialogue, KGs are particularly valuable as they are capable of representing complex relationships that are frequently present in dialogue.
However, applying IE techniques to dialogue presents unique challenges not found in static documents. There is a fundamental difference between the fragmented, context-dependent nature of human conversation and the organized format of curated texts. Dialogue is characterized by multiple speakers, frequent use of pronouns and anaphora, implicit knowledge, frequent topic changes, and noise. These features complicate the IE process, requiring models that can not only understand individual utterances but also track entities, relations, and context across multiple turns and speakers.
While as far as we are aware, there are no existing published surveys on knowledge graph construction or information extraction as a whole from dialogue, there do exists works synthesizing previous applications of Relation Extraction (RE) to dialogue. Zhao et. al. [47] have published a survey on RE that dedicates space to dialogue as a target domain, and Xu et. al. [41] have released a survey focused on information extraction using LLMs, with a dedicated section on Dialogue Relation Extraction (DialogRE). Additionally, there are surveys on dialogue state tracking (DST) [13, 2]. While DST is distinct from RE, traditionally following a pipeline of intent classification and slot filling using a strict schema as seen in Figure 3, it is an information extraction task that is consistently used in dialogue systems. An emerging line of research in DST entails the generation of flexible schemas and knowledge graphs as dynamic state histories.
This survey aims to synthesize and build upon these existing works, providing a focused overview of the current state of research in extracting relational triples and knowledge graphs from conversational data. Primarily, we focus on the following three tasks:
• Relation Extraction (RE): The task of identifying and extracting potential n-ary relations from an unstructured source. We specifically consider binary relations in the form (subject, predicate, object). Additionally, we consider tasks such as personal attribute and sentiment extraction where the output takes the form of semantic triples.
• Dialogue State Tracking (DST): The task of tracking the state of a dialogue, often represented as a set of slot-value pairs defined by a schema.
• Knowledge Graph (KG) Construction: The end-to-end process of building a knowledge graph from unstructured text, through knowledge acquisition, refinement, and evolution [48].
Links:
- Final paper: https://drive.google.com/file/d/1HJoKXvM40D5s8DO22B__68f9fKVcmTAn/
- Final presentation (slides): https://drive.google.com/file/d/1hgnO6a0FjtHTnhBq3EHiGm5wd71g59kL/
- Final presentation (video): https://youtu.be/QEIKy4BtLQk
