Personal Attribute Prediction from Conversations

Abstract

Personal knowledge bases (PKBs) are critical to many applications, such as Web-based chatbots and personalized recommendation. Conversations containing rich personal knowledge can be regarded as a main source to populate the PKB. Given a user, a user attribute, and user utterances from a conversational system, we aim to predict the personal attribute value for the user, which is helpful for the enrichment of PKBs. However, there are three issues existing in previous studies: (1) manually labeled utterances are required for model training; (2) personal attribute knowledge embedded in both utterances and external resources is underutilized; (3) the performance on predicting some difficult personal attributes is unsatisfactory. In this paper, we propose a framework DSCGN based on the pre-trained language model with a noise-robust loss function to predict personal attributes from conversations without requiring any labeled utterances. We yield two categories of supervision, i.e., document-level supervision via a distant supervision strategy and contextualized word-level supervision via a label guessing method, by mining the personal attribute knowledge embedded in both unlabeled utterances and external resources to fine-tune the language model. Extensive experiments over two real-world data sets (i.e., a profession data set and a hobby data set) show our framework obtains the best performance compared with all the twelve baselines in terms of nDCG and MRR. CCS CONCEPTS• Computing methodologies → Information extraction.

A comprehensive survey of personal knowledge graphs

“…Populating a PKG from conversational utterances can sometimes be a challenging task without the presence of a labeled dataset of utterances for training a language model for this task. The DSCGN architecture (Y. Liu, Chen, & Shen, 2022) overcomes this hurdle by deploying a distant supervision strategy for document‐level supervision over external sources like Wikipedia pages, and a contextualized word‐level supervision using a label guessing technique over unlabeled data. This model first carries out distant supervision by obtaining Wikipedia articles of personal attribute values (e.g., scientist, actor, teacher) and related topics to form the document‐level supervision where is some Wikipedia page having an attribute value .…”

“…Another method to tackle the lack of labeled utterance data for personal attribute extraction is a framework called PEARL (Y. Liu, Chen, Shen, & Chen, 2022) that predicts such user attributes by combining the attribute value representations with the biterm set generated for each attribute value. In particular, the attribute value representation is fused with biterm semantic information using Gibb's sampling process in the framework of biterm topic model (Yan et al, 2013).…”

Low-Resource Personal Attribute Prediction from Conversations

“…Recently, some neural network based models have been explored for this task. These models resort to labeled utterances (Tigunova et al 2019), external data (e.g., Wikipedia and Web pages) (Liu, Chen, and Shen 2022) or both (Tigunova et al 2020) as resources of training data. However, there exist three issues in these previous works: (1) they rely on many resources of training data but these resources are not always available and expensive to fetch, which limits their adaptability to new domains or new data; (2) the attribute knowledge embedded in the unlabeled utterances is underutilized; (3) their performance over some difficult personal attributes (e.g., profession and hobby) is still unsatisfactory.…”

“…Performance on the task of personal attribute prediction from conversations. All the results of the personal attribute prediction baselines are taken from DSCGN(Liu, Chen, and Shen 2022). The performance of all the weakly supervised text classification methods is reproduced via their open-source solutions.…”