The information overload is a well-known problem that impacts the digital experience of users when they need to find interesting items in a large set of possible options [1]. That is the case of looking for a book to read, a digital camera to buy, a movie to watch, etc. Another scenario when this problem is particularly felt is when a user is looking for a new job. In this context the only strategy people can adopt is to manually selecting interesting job calls, thus resulting in a clumsy user experience. In the literature, this problem has been investigated in the information-retrieval (IR) and recommender-system (RS) research areas. In the first case, systems do not take into account the user past preferences and retrieve the most relevant items according to the user query. In the second case a user profile is built, allowing to provide a personalized ranked list of items.
The GUapp platform has been designed to find and discover job positions among job offers in the Italian public administration 1 . The first version of GUapp's RS classified jobs according to their topics, obtained by means of the Latent Dirichlet Allocation (LDA) algorithm, and a k-Nearest Neighbor (k-NN) method was used to compute a personalized recommendation list [2] 2 . Moreover, GUapp offers a natural-language based interaction through a chatbot. The GUapp's chatbot allows the user to define her interests, describe her skills, and filter out results that did not match with her requirements. Although the first version of Guapp was undoubtedly effective from the recommendation perspective, there were a few limitations in terms of User Experience. The combination of k-NN and LDA, in fact, does not allow the system to explain to the user why an item was recommended. This because the latent features learned by the model are not easily interpretable. Hence, the dialogues were modeled as a filtering procedure for the recommendations list without interacting neither exchanging information with the RS itself. Users that converse with the system can only provide preferences from a limited set of features that are foreseen a priori, that are city and profession.
To beat these issues, we opted for a new Conversational Agent that converses with users exploiting a domain-specific Knowledge Graph (KG). The GUapp KG is obtained by merging some sub-graphs from state-of-the-art solutions like Dbpedia3 and new triples scraped from external sources such as the ISTAT 4 website. On this line, conversations can reach a fine grained level of details regarding job aspects to recommend to users and enhance its expressiveness.
The work proposed in this paper cuts across two principal research areas: recommender systems and conversational agents. Recommender Systems (RSs) can in turn be grouped into two main classes: Collaborative Filtering (CF) and Content-based (CB). To support the user during the decision making process, CF RS exploits the user community in order to identify interesting items for a given individual, following the intuition that similar users like similar items. Conversely, CB RSs are based on matching user preferences with items textual content [1].
GUapp implements a CB RS since its goal is to match the user requests/preferences with the textual description of the job call. The specific task of job recommendation is widely investigated in the literature [3]. Malinowski et al. [4] build a multi-slot profile with information about demographic data, job experience and IT skills. Similarly, in GUapp, this kind of information was first acquired from the LinkedIn social network. This strategy, useful for addressing the coldstart problem, is now implemented by means of conversations driven by a KG. Recommending items by means of a set of rules which verify whether the user tastes are satisfied or not was dealt at first by the Knowledge based RSs [5]. Here the RS performs reasoning on ontologies and Knowledge Bases (KBs) to find items that match the user needs. Carrer-Neto et al. [6] recommend movies that belong to the same ontological classes of the items that the user liked, and the final list is further processed by means of a social context derived from Social Networks in which the user is signed-up. Tarus et al. [7] exploit both ontological and CF features to compute the learning resources to recommend for some learner targets.
On the other hand, Conversational Agents (CAs) are software agents able to interact with the user through natural language. They can be classified in two main classes: end-to-end and modular systems [8]. The former are generally based on neural networks [9,10,11,12] and they learn a dialog model from a set of past conversations [13]. The modular systems are pipeline-based agents composed of a set of modules, each with a specific function [14,15,16,17].
The main difference between a Conversational Recommender System (CoRS) and a traditional RS is the interaction with the user that is more efficient and natural [18]. Accordingly, the construction of the user profile is incremental and the user can express her preferences at different stages of the recommendation process by an interactive human-like dialog [19]. The aforementioned task is well suited to Knowledge-based and KG-based RSs. An example can be found in [20], where a comprehensive KG is built upon a specific domain to lead dialogues and recommendations. Recently, CoRSs have been effectively used and tested in different domains such us music, movie, and book [21,22]. However, to the best of our knowledge, GUapp represents the first attempt of implementing a CoRS for the job-recommendation task.
In Figure 1 the initial GUapp's architecture is sketched together with the new one, with six main components: the Orchestrator, the Chatbot, the Recommender System, the User Profiler, the Crawler and the Knowledge Graph. The Orchestrator manages the interaction between the different components of the system. For example, it invokes the recommender system when user asks for receiving a list of job positions based on the information stored in her profile.
The Chatbot is the component of GUapp which allows user to interact through natural language. It is powered by DialogFlow 5 , a Google platform for designing and integrating conversational user interfaces, and performs two main tasks: intent and entity recognition. The Intent Recognizer (IR) analyzes the request of the user expressed in natural language and understands her goal. E.g. if the user writes "I am looking for a new job in Bari" the IR identifies as intent new_job. The GUapp's Chatbot adopts a mixed initiative interaction, allowing more fine grained conversations. For instance, the user may add her skills and professional titles through the profile intent. She may refine also her favourite job locations or occupations with the refine_location and refine_occupation intents respectively. The system can also take the initiative and ask for unspoken user features, that could eventually be inferred by the KG, and engage a negotiation about job postings that do not have a full match with the user requirements. In details, if the user profile is a tuple (๐ต๐๐๐, ๐ถ๐๐๐๐ข๐ก๐๐๐ธ๐๐๐๐๐๐๐, ๐ โ๐ท) and the retrieved results do not have a complete match, the system may ask: "What do you think about Computer Science in Bari?" or "Is it still good for you a position in Lecce?". Once the intent has been recognized, the Entity Recognizer (ER) is invoked. ER is implemented through DialogFlow as well, and it is powered by the KG entities. It adopts a fuzzy-matching strategy in order to identify real-world entities in the user sentence. In the case the match succeeds, the recognized entity is returned. In the example above, the entity is Bari, a city of southern Italy.
The User Profiler (UP) instead stores and updates information about user preferences and habits. Differences here with the first version of GUapp are on the information type stored in UP. We now collect data about all the user preferences provided during conversations instead of users' likes. Indeed, user can refine the retrieved list of relevant job positions as described above. Due to some variations on Microsoft's terms of use, the new version of GUapp does no more have access to the LinkedIn6 data. At first we found useful to collect information stored in the social network to profile users by simply asking them a login with their LinkedIn credentials. This was an effective strategy for addressing the cold-start problem. Now, when the user signs in to the app for the first time, her profile is empty (i.e. cold-start situation). Thanks to the new chatbot powered by our KG, the user talks with GUapp about her skills, wishes and possibilities, then the system is able to rank job calls by exploiting the information that the user provided.
The Recommender System is another core component of GUapp. It exploits an Elasticsearch7 index, which stores all the information scraped by the Crawler enriched with all the linked entities of our KG. We are now stressing the choice of a similarity function that effectively returns a ranked list of item based on the data collected by the UP. The main intuition behind adopting this model is the possibility to make dialogues as interactive and efficient as possible, allowing the system to negotiate with users about results that do not completely match with their preferences. The UP is actively updated to the new user inputs, guaranteeing the recommendations to be adaptive. As a matter of fact, GUapp implements a Crawler that daily extracts the job positions from Gazzetta Ufficiale, the official journal of record of the Italian government. Finally, the system has been equipped with a domain-specific Knowledge Graph built upon several sources, like Dbpedia and the ISTAT website as stated before. The latter identifies a collection of raw data that are not in form of KG. This leads us to model a new Ontology on which the overall KG can rely. The GUapp Ontology in fact defines relations and hierarchies about professions, domain of competences, locations and so forth wisely integrating the raw information with the already structured ones. To make the system more specialized on the job-opening domain, we plan to enrich our KG of further facts related to the job recommendation task. This KG allows the system to have natural dialogues with users about the discussed domain and the quality of recommendations results improved as well.
In this paper we presented a platform for searching jobs in the Italian public administration and the main ongoing work for improving its performances and the user experience. GUapp consists of a recommender system that suggests relevant jobs to the users. One of the most interesting aspects of the new version is the integration of a KG that helps driving the dialog, making the interaction more natural. Indeed, the preference elicitation becomes incremental, with the possibility of refining and improving the user requests. We are planning of introducing several fine-grained features for making the ecosystem competitive with other state-of-the-art solutions, as well testing the User Experience with an A/B test.